Stochastic Block Model using Pyro

[aladinD]

@fehiepsi

It is constant during sampling process (you can look at the summary statistics in that comment), so taking the last value is fine IMO.

Exactly, this was my notion, too! @cosmicc

@cosmicc

The Model can also be easily extended to a mixed membership model by using a second categorical distribution and exchanging one of the vectors with the corresponding 1-hot encoded vector in the matrix multiplication.

Please, feel free to share your working model. I think that this might also be interesting for some other folks who’d like to do SBM (and its different variants) using Pyro

[cosmicc]

@aladinD @fehiepsi .
What is the intuition behind expecting the latent states to be constant?
Wouldn’t there always be a remaining probability of jumping back into an lower probable state, especially if there is an overlap between the pdfs?
Cheers

[cosmicc]

interestingly the SBM is kind of self regulatory. Though I iteratively increased the input block number till 11 it never surpassed 7. After that the posterior beta distributions are smashed on the LHS of the chart